JP7440938B2 - Event prediction system, event prediction method and program - Google Patents

Description

The present invention relates to event prediction technology based on time-series sensor data.

In recent years, in the manufacturing industry, there has been a push to make manufacturing plants smarter. By constantly monitoring the operating status of the production line using a large number of sensors, and accumulating and analyzing the status as time-series data, it is possible to detect abnormalities in equipment (Non-patent Documents 25, 32) and quality control (Non-patent Documents 14) Efforts are being made to improve productivity from all aspects. An important issue common to these efforts is the effective acquisition of knowledge from the large-scale data collected and the development of future prediction technology based on this knowledge. In particular, time-series data obtained from manufacturing plants is complex data with multiple domains (equipment, sensors, time, etc.) and often has multifaceted patterns. On a production line, not only the time transition of multiple work processes (patterns), but also common/different patterns are created for each work line due to parallel work on multiple lines. In order to effectively understand the causes of defective products and equipment failures, it is necessary to flexibly express such multifaceted and dynamic patterns and at the same time clarify the hidden causal relationships between them.

In addition, for tasks envisioned in smart factories, knowing in advance the occurrence of various events, such as breakdowns, defects, and reductions in processing accuracy, will expand the options for countermeasures. In other words, it is desirable that future prediction technology for large-scale sensor data has a longer-term prediction ability (Non-Patent Document 15).

Research on analysis of sensor data is progressing in various fields such as databases and data mining (Non-Patent Documents 2, 17, 19, 22, 24, 25). Auto regressive models (AR) and linear dynamical systems (LDS) are representative technologies, and there are many sensor data analysis and prediction methods based on these (Non-patent Document 13). .

Regime-Cast (Non-Patent Document 15) has the ability to estimate nonlinear dynamic systems in real time from multidimensional sensor data that continues to be generated in large quantities, and to continue to adaptively predict the future. However, although this method uses sensor streams as input and shows high performance in predicting actual measured values of sensor data, it does not support prediction of event data such as normality/abnormality.

Furthermore, pattern discovery and clustering for time-series big data are also important issues (Non-Patent Documents 8, 10, 11, 16, 28, 29, 31). Matsubara et al. (Non-Patent Document 18) proposed TriMine as a method for analyzing large-scale event tensors. TriMine classifies given data into multiple topics and detects potential trends and patterns, but it targets discrete event data like click logs on the web and time series like IOT sensor data. It is not possible to express dynamic patterns of sequences or their groups (regimes), and the problems they deal with are different. Additionally, TriMine has no ability to predict events.

Research on analysis of nonlinear dynamic characteristics based on Deep Neural Networks is also active (Non-Patent Documents 3, 9, 26, 27). In Non-Patent Document 21, Qin et al. proposed a method for predicting stock prices with high accuracy by modeling important dimensions in the input time series and important dimensions in the special space after dimension reduction over two layers. . On the other hand, in the task of predicting events that occur discontinuously, as in this study, the mainstream is a method of modeling the intensity of event occurrence (Non-Patent Documents 5, 6, 20, 30). For example, RMTPP (Non-Patent Document 5) proposes a nonlinear model for predicting the time and type of the next event based on past event history. However, these methods target categorical data consisting only of event history, and cannot predict events using continuous data consisting of actual measured values from sensors.

C. M. Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer, 2006. G. E. Box, G. M. Jenkins, and G. C. Reinsel. Time Series Analysis: Forecasting and Control. Prentice Hall, Englewood Cliffs, NJ, 3rd edition, 1994. P. Chen, S. Liu, C. Shi, B. Hooi, B. Wang, and X. Cheng. Neucast: Seasonal neural forecast of power grid time series. In IJCAI, pages 3315-3321, 2018. K. Cho, B. van Merrienboer, D. Bahdanau, and Y. Bengio. On the Properties of Neural Machine Translation: Encoder-Decoder Approaches. arXiv e-prints, page arXiv:1409.1259, Sep 2014. N. Du, H. Dai, R. Trivedi, U. Upadhyay, M. Gomez-Rodriguez, and L. Song. Recurrent marked temporal point processes: Embedding event history to vector. In KDD, pages 1555-1564, 2016. N. Du, Y. Wang, N. He, and L. Song. Time-sensitive recommendation from recurrent user activities. In NIPS, pages 3492-3500, 2015. J. G. DAVID FORNEY. The viterbi algorithm. In Proceedings of the IEEE, pages 268-278, 1973. D. Hallac, S. Vare, S. Boyd, and J. Leskovec. Toeplitz inverse covariance-based clustering of multivariate time series data. In KDD, pages 215-223, 2017. S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, Nov. 1997. T. Honda, Y. Matsubara, R. Neyama, M. Abe, and Y. Sakurai. Multi-aspect mining of complex sensor sequences. In ICDM, 2019. K. Kawabata, Y. Matsubara, and Y. Sakurai. Automatic sequential pattern mining in data streams. In CIKM, pages 1733-1742, 2019. D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. L. Li, J. McCann, N. Pollard, and C. Faloutsos. Dynammo: Mining and summarization of coevolving sequences with missing values. In KDD, 2009. Y. Li, J. Wang, J. Ye, and C. K. Reddy. A multi-task learning formulation for survival analysis. In KDD, pages 1715-1724, 2016. Y. Matsubara and Y. Sakurai. Regime shifts in streams: Realtime forecasting of co-evolving time sequences. In KDD, 2016. Y. Matsubara, Y. Sakurai, and C. Faloutsos. Autoplait: Automatic mining of co-evolving time sequences. In SIGMOD, pages 193-204, 2014. Y. Matsubara, Y. Sakurai, and C. Faloutsos. The web as a jungle: Non-linear dynamical systems for co-evolving online activities. In WWW, pages 721-731, 2015. Y. Matsubara, Y. Sakurai, C. Faloutsos, T. Iwata, and M. Yoshikawa. Fast mining and forecasting of complex timestamped events. In KDD, pages 271-279, 2012. Y. Matsubara, Y. Sakurai, B. A. Prakash, L. Li, and C. Faloutsos. Rise and fall patterns of information diffusion: model and implications. In KDD, pages 6-14, 2012. H. Mei and J. Eisner. The neural hawkes process: A self-modulating neutral multivariate point process. In NIPS, pages 6757-6767, 2017. Y. Qin, D. Song, H. Chen, W. Cheng, G. Jiang, and G. W. Cottrell. A dual-stage attention-based recurrent neural network for time series prediction. In IJCAI, pages 2627-2633, 2017. T. Rakthanmanon, B. J. L. Campana, A. Mueen, G. E. A. P. A. Batista, M. B. Westover, Q. Zhu, J. Zakaria, and E. J. Keogh. Searching and mining trillions of time series subsequences under dynamic time warping. In KDD, pages 262-270, 2012. J. Rissanen. A Universal Prior for Integers and Estimation by Minimum Description Length. Ann. of Statist, 11(2):416-431, 1983. Y. Sakurai, Y. Matsubara, and C. Faloutsos. Mining and forecasting of big time-series data. In SIGMOD, pages 919-922, 2015. Y. Sakurai, S. Papadimitriou, and C. Faloutsos. Braid: Stream mining through group lag correlations. In SIGMOD, pages 599-610, 2005. I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence learning with neural networks. In NIPS, pages 3104-3112. 2014. Tsungnan Lin, B. G. Horne, P. Tino, and C. L. Giles. Learning long-term dependencies in narx recurrent neural networks. IEEE Transactions on Neural Networks, 7(6):1329-1338, 1996. P. Wang, H. Wang, and W. Wang. Finding semantics in time series. In SIGMOD Conference, pages 385-396, 2011. S. Wang, K. Kam, C. Xiao, S. R. Bowen, and W. A. Chaovalitwongse. An efficient time series subsequence pattern mining and prediction framework with an application to respiratory motion prediction. In AAAI, pages 2159-2165, 2016. S. Xiao, J. Yan, X. Yang, H. Zha, and S. Chu. Modeling the intensity function of point process via recurrent neural networks, 2017. R. Zhao and Q. Ji. An adversarial hierarchical hidden markov model for human pose modeling and generation. In AAAI, 2018. Y. Zhou, H. Zou, R. Arghandeh, W. Gu, and C. J. Spanos. Non-parametric outliers detection in multiple time series A case study: Power grid data analysis. In AAAI, 2018.

As explained above, conventional event prediction methods and systems have been proposed that target time-series tensor data and perform event prediction using characteristic patterns of time-series data without requiring prior knowledge of time-series patterns. It has not been.

The present invention has been made in view of the above, and provides an event prediction system, method, and program for the event prediction system that targets time-series tensor data and enables long-term and highly accurate event prediction through data summarization processing. This is what we provide.

The event prediction system according to the present invention continuously extracts model parameters of multifaceted dynamic patterns from time-series sensor data continuously collected from multiple types of sensors placed on multiple observation targets. a first feature extracting means; a second feature extracting means for sequentially converting the time-series sensor data into features using the model parameters into summary information including modeling information and its error information; and the summary information and a prediction means for outputting the probability of occurrence of a predetermined event in a predetermined time period.

Further, in the event prediction method according to the present invention, the first feature extracting unit of the computer continuously collects data from multiple types of sensors placed on multiple observation targets, and stores the time series in the storage unit. Model parameters of a multifaceted dynamic pattern are continuously extracted from sensor data and stored in the storage unit, and a second feature extracting unit of the computer extracts the model parameters and the time-series sensor data from the sensor data. The time-series sensor data is read out from the storage unit and sequentially converted into feature quantities into summary information including modeling information and its error information, and stored in the storage unit, and the prediction unit of the computer converts the summary information into the storage unit. It is read from the section and used as input, and outputs the probability of occurrence of a predetermined event in a predetermined time period.

Further, the program according to the present invention continuously extracts model parameters of multifaceted dynamic patterns from time-series sensor data continuously collected from multiple types of sensors placed on multiple observation targets. a first feature extracting means; a second feature extracting means for sequentially converting the time-series sensor data into summary information including modeling information and its error information using the model parameters; The computer functions as a prediction means that outputs as input the probability of occurrence of a predetermined event in a predetermined time period.

According to these inventions, time-series sensor data is continuously collected from multiple types of sensors placed on multiple observation targets, and model parameters of multifaceted dynamic patterns are determined from the collected time-series sensor data. Extraction is continuously performed by the first feature extraction means. Next, the time-series sensor data is sequentially converted into features by the second feature extracting means using the model parameters into summary information including modeling information and its error information. Then, the prediction means inputs the summary information and outputs the probability of occurrence of a predetermined event in a predetermined time period. Therefore, there is no need for prior knowledge of time-series patterns contained in time-series sensor data, and changing points and potential behavior of patterns (regimes) can be determined based on, for example, time transitions and multiple perspectives between observation targets. be understood. Furthermore, it becomes possible to discover characteristic patterns in large-scale time-series sensor data and use them to predict events over a long period of time. Note that the sensors may be placed directly on the observation target, or may be installed in such a manner that the observation target can be observed remotely.

According to the present invention, by extracting and summarizing feature quantities from multiple angles from time-series sensor data, it is possible to perform long-term and highly accurate event prediction with a simple configuration.

1 is an overall block diagram showing an embodiment of an event prediction system according to the present invention. A diagram showing an example of the processing status of information imported from smart factory data to which the present invention is applied, in which (a) shows original sensor data, and (b) shows pattern detection results from the original data. , (c), and (d) are diagrams showing typical regime examples of a case (d) in which an emergency stop is made after a predetermined time and a case (c) in which it is not, based on original data. 1 is a diagram showing an outline of a proposed model according to the present invention. FIG. 3 is a transition diagram for explaining the basic concept of the proposed algorithm according to the present invention. FIG. 7 is a comparison diagram of accuracy when changing the number of steps to predict, _ls . FIG. 4 is a comparison diagram of prediction accuracy when changing the window width of the mini-batch used during network learning. FIG. 3 is a comparison diagram showing precision and recall of prediction results. FIG. 3 is a diagram showing changes in prediction accuracy of the present prediction system with respect to the number m of detected segments. FIG. 3 is a diagram showing the relationship between the number of learning samples and prediction accuracy. FIG. 7 is a diagram showing the calculation cost of the present prediction system when the number of equipment w, the number of sensors d, and the sequence length n are changed.

The present invention relates to an event prediction method, preferably for large-scale time-series sensor data. As an example, the present invention comprehensively analyzes and summarizes multifaceted time-series patterns based on multiple viewpoints from factory equipment sensor data, which is composed of the triplet of (facility, sensor, time). Regarding technology for long-term event prediction. More specifically, when given time series data consisting of actual measured values of sensor data such as rotation speed, operating voltage, and equipment temperature for each piece of equipment installed in a factory, (a) basic time series By extracting patterns, common patterns between each piece of equipment, and equipment-specific patterns, and statistically summarizing them, (b) predict future events. Furthermore, these processes are (c) linear with respect to data size. As described later, in experiments using real data, we confirmed that this prediction method captures characteristic time-series patterns included in sensor data of factory equipment from multiple angles and predicts long-term events. Furthermore, as described below, compared with the latest existing method (comparative example), it was revealed that significant accuracy and performance improvements were achieved.

In other words, this prediction system analyzes the number and change points of typical patterns (hereinafter referred to as regimes) included in time-series data from multiple angles, and accurately grasps the operating status of the system to predict future occurrences. Predict events. More specifically, when given large-scale time-series sensor data collected from multiple sensors at multiple locations of equipment, events after a predetermined time, ie, 1 _s steps ahead, are predicted.

More specifically, (a) detecting multifaceted patterns and their changing points from sensor data and summarizing them as summary information will (b) help realize long-term and highly accurate predictions. Furthermore, (c) these processes are performed at high speed.

The present invention will be described below with reference to the drawings. FIG. 1 shows an overall block diagram of an event prediction system (hereinafter referred to as prediction system 1) according to the present invention. This prediction system 1 transmits large-scale time-series sensor data related to operating conditions from respective sensor groups 21 installed in observation targets 20, ..., which are a plurality of facilities such as factories, through wired or wireless communication channels. and a computer having a control unit 10 consisting of a processor (CPU) that extracts feature quantities from each imported time series data and further executes event prediction processing after a predetermined period of time. . Furthermore, in this embodiment, machine learning is used, and parameters applied to prediction processing are updated through machine learning. Details of FIG. 1 will be described later.

First, in order to understand the prediction process, a specific example shown in FIG. 2 will be explained. FIG. 2 shows sensor data from a smart factory as an example of the observation target 20 (FIG. 1), and shows information used (input) for prediction processing. FIG. 2(a) shows the original sensor data, in which three sensor values (rotation speed :Speed, operating voltage: Load, equipment temperature: Temp). In FIG. 2(a), a portion filled with a black rectangle indicates that the corresponding equipment is in an emergency stop. Note that the waveform of the operating voltage: Load in FIG. 2(a) roughly overlaps with the waveform of the rotational speed: Speed. FIG. 2(b) shows the result of pattern extraction from the original data by this prediction system. Vertical lines in FIG. 2(b) indicate times when the time-series pattern changes, and segments belonging to the same regime are represented in the same shaded color. By simultaneously analyzing time-series data obtained from multiple pieces of equipment, the prediction system 1 analyzes multifaceted patterns, that is, not only temporal transitions of patterns within each piece of equipment, but also patterns that are common or different between pieces of equipment. It is possible to detect.

FIGS. 2(c) and 2(d) show typical examples from the original data of cases in which an emergency stop occurs after l _s = 200 steps (approximately 17 minutes) and cases in which it does not. The left side of FIGS. 2(c) and 2(d) shows the segmentation results. θ ₁ to θ ₅ on the right represent common time series patterns (ie, regimes), and their transitions are visualized. The value of p200 is the emergency stop probability 200 steps ahead that is output by this prediction system when the partial sequence and its pattern detection results corresponding to the left diagrams in FIGS. 2(c) and 2(d) are given. In the right-hand diagrams of FIGS. 2(c) and 2(d), thick arrows are displayed between regimes where more transitions are detected. Furthermore, the size of the circle indicates the size of the period during which the regime occurred. Looking at FIG. 2(d), the rotational speed Speed increases (θ ₅ ) before the equipment makes an emergency stop, and this tendency is expressed by the appearance of a transition between regimes θ ₄ and θ ₅ . In fact, this prediction system 1 accurately predicts emergency stops, and p200 shows a high value. In other words, by detecting latent patterns contained in the data, not only can the process leading to an emergency stop be analyzed from multiple angles, but also long-term and highly accurate predictions can be made using the summary information. In addition, in FIG. 2(c), transitions such as regimes θ ₂ , θ ₃ , θ ₂ , θ ₁ , and θ ₂ without any sign of emergency stop are observed, and p200 also shows a low value.

As an example of factory equipment sensor data handled by this prediction system 1, three types of sensor data from 55 equipment operated at Mitsubishi Heavy Industries Engine & Turbocharger Corporation on October 1, 2017 are shown. This data is expressed as a triplet (facility, sensor, time), each consisting of w pieces of equipment, d types of sensors, and n periods (for example, in units of 5 seconds). Such sensor data can be expressed as a third-floor tensor X ∈ R ^{w × d × n} , where the element x _ij (t) of the tensor Show value. In this embodiment, such sensor data is referred to as a multidimensional time series tensor.

This prediction system 1 predicts an equipment alert _ls steps ahead from a given time series tensor X, and the necessary processing for this is shown below.

That is, when the time series tensor X(t _s :t _e ) is given, the alert label y(t _e +l _s ) for l _s steps ahead is predicted based on the following equation (1).

у(t _e +l _s ) ≒F(X(t _s :t _e )) (1)
Note that t _s : _te represents the window of the sequence used for prediction (a predetermined period from the current moment to the past), and F is the proposed model.

Here, in order to predict the alert label y(t _e +l _s ) with high accuracy, we build a model based on a probabilistic model and deep learning, and use the given sensor data to predict, for example, the cause of a failure (alert). Extract high-dimensional and nonlinear dynamic characteristics. Specifically, this prediction system 1 executes the following three processes (P1), (P2), and (P3).

(P1) Multifaceted detection of latent dynamic patterns
(P2) Feature extraction based on dynamic patterns
(P3) l Long-term prediction for _s steps ahead First, the relationship with FIG. 1 will be explained for each process (P1), (P2), and (P3). In FIG. 1, the control unit 10 is connected to a storage unit 100, a display unit 121 that displays a window to be described later, for example, and an operation unit 122 that accepts instructions from the outside. The storage unit 100 includes a control program storage unit 101, a data stream storage unit 102 that stores time-series sensor data input from each sensor group 21, and a neural network that constitutes artificial intelligence (AI) that is applied to prediction processing. It includes a parameter storage unit 103 that stores model parameters (weights of each edge, etc.). The control program storage unit 101 stores program data and various necessary arithmetic expression data for executing event prediction processing, which will be described later. In addition to the data stream storage unit 102, the storage unit 100 also stores various processes such as “(P1) Multifaceted detection of latent dynamic patterns” and “(P2) Feature extraction based on dynamic patterns” which will be described later. It has a work area (storage unit) that temporarily stores each data obtained during the execution of "(P3) long-term prediction of the _next step".

The control unit 10 functions as a data acquisition processing unit 11, a feature quantity extraction unit 12, a prediction unit 13, and a parameter updating unit 14 by executing a control program. The data acquisition processing unit 11 acquires time-series sensor data from the sensor group 21 of each observation target 20 (each piece of equipment in the factory) via the network 110.

The feature extraction unit 12 executes the processes "(P1) Multifaceted detection of latent dynamic patterns" and "(P2) Feature extraction based on dynamic patterns", which will be described later. The prediction unit 13 executes the process “(P3) long-term prediction for _s steps ahead”. In this embodiment, the prediction unit 13 performs prediction processing by applying parameters from the parameter storage unit 103. Details of each process will be described later.

The machine learning device 30 includes a control section 300 and a storage section 310 that are made up of a computer with a built-in processor, as well as a display section 321 and an operation section 322. The storage unit 310 includes a learning program storage unit 311, a data stream storage unit 312, and a parameter storage unit 313. The data stream storage unit 312 captures time-series sensor data input from each sensor group 21 through communication or via an external memory, or captures and stores data once written in the data stream storage unit 102.

The control unit 300 functions as a data acquisition processing unit 301, a feature extraction unit 302, and a machine learning unit 303 by executing a learning program from a learning program storage unit 311. The data acquisition processing unit 301 is similar to the data acquisition processing unit 11, and can further automatically or manually set the acquisition period of the acquired data (for example, for the most recent one week) as appropriate. The feature extraction unit 302 is provided as necessary, and checks the processing by adjusting the conditions of the processing (P1) and (P2) as appropriate, for example, in response to changes in factory equipment or other changes in the situation.

The machine learning unit 303 preferably performs machine learning by applying, for example, “supervised learning” to the time-series sensor data for the most recent predetermined period, and stores the parameters that are the learning results in the parameter storage unit 313. Then, the parameter storage unit 103 is updated as necessary via the parameter update unit 14 or in response to an instruction from the operation unit 322 of the machine learning device 30. Note that, in addition to the form of the separate machine learning device 30, various forms of machine learning can be adopted. For example, the input data may be extracted for a predetermined period from the data stream storage unit 102. Alternatively, learning may be performed using the prediction unit 13 during a system stop period (for example, at night), and the parameters that are the learning results may be updated.

Next, the outline and necessary definitions of the "proposed model" are shown in Table 1.

<Proposed model>
(P1) Detection of latent dynamic patterns When a multidimensional time series tensor X _is given, this prediction system first divides X into a set of m segments S={s ₁ ,. Divide it into parts to understand its characteristics. s _i is composed of the start point t _s , end point t _e , and facility number of the i-th segment (i.e., s _i ={t _s ,t _e ,facilityID}), and each segment is assumed to have no overlap. Then, the discovered segment set is classified into groups of similar segments. In this prediction system, these groups are called "regimes."

・Definition 1 (regime)
Let r be the optimal number of segment groups. Each segment s is assigned to one of the segment groups. Furthermore, to express the regime to which each segment belongs, segment membership is newly defined.

・Definition 2 (segment membership)
When a multidimensional time series _tensor X is given, let F={f ₁ , _. 1≦f _i ≦r).

Thereby, the multidimensional time series tensor X can be expressed as {m, r, S, Θ, F} using m segments and r regimes. Next, the present prediction system statistically models the multidimensional time series tensor X based on the obtained regime information and extracts important features.

(P2) Feature extraction based on dynamic patterns Each regime is expressed as a statistical model Θ={θ ₁ , . . . , θ _r ,Δ _r×r }. In this study, we use Hidden Markov Model (HMM) to express the behavior of multidimensional time series tensor X. HMM is a type of probabilistic model that assumes a Markov process with hidden states, and is widely used as a time series processing method in various fields including speech recognition. HMM is a triplet of initial probability Π={π _i } ^k _i=1 , transition probability A={a _ij } ^k _i,j=1 , and output probability B={b _i (x)} ^k _i=1 (i.e., θ={Π,A,B}). Here, k indicates the number of latent states of the HMM. In this prediction system, it is assumed that the output probability B is generated from a multidimensional Gaussian distribution. This represents a sequence of multidimensional vectors as a probabilistic model (ie, B~{N(μ _i ,σ ² _i )} ^k _i=1 ). When the HMM model parameters θ={Π,A,B} and a certain user sequence х are given as input data, the likelihood P(х|θ) of х is expressed as the following equation (Equation 1): calculated.

Here, p _i (t) indicates the maximum probability of latent state i at time t, and n is the sequence length of х. This likelihood is calculated using the Viterbi algorithm (non-patent document 7), which is a type of dynamic programming, based on the transition diagram shown in FIG. Here, we further introduce the regime transition matrix Δ _r×r as a new concept.

・Definition 3 (regime transition matrix)
Δ _r×r is called the transition matrix of r regime groups. Here, the element δ _ij ∈Δ indicates the transition probability from the i-th regime to the j-th regime. That is, the condition is 0≦δ _ij ≦1, Σ _j δ _ij = 1. Using the above model, the multidimensional time series tensor Realize.

・Definition 4 (latent state tensor)
The latent state sequence Z={Z ₁ ,. . . ,Z _w } of the HMM for each equipment is called the latent state tensor. Here, Z _i ={z _ij (1),. . . ,z _ij (n)} ^d _j=1 , and z _ij (t) is the average of the data set It consists of a set of variances {μ,σ}.

・Definition 5 (error tensor)
The error ε={E ₁ , . . . , E _w } when the multidimensional time series tensor X is modeled by the latent state tensor Z is called an error tensor. In this prediction system, it is assumed that the output probability B of the HMM follows a multidimensional Gaussian distribution, so the error e _ij (t) ∈ E _i at time t in the j-th sensor of the i-th equipment is It is expressed as (Equation 2).

That is, the time series tensor X is transformed into a latent state tensor Z and an error tensor ε such that and extract important features. Here, IGPDF (Inverse Gaussian Probability Density Function) represents the inverse function of the probability density function in Gaussian distribution.

(P3) Long-term prediction for ls steps ahead In conclusion, the above equation (1) can be rewritten as the following equation (2).

Here, F represents a prediction model. That is, when a time series tensor X is given, the proposed method extracts important features by summarizing X with a latent state tensor Z and an error tensor ε, applies the proposed model F to them, _and Make long-term predictions with high accuracy.

<Algorithm for processing (P1), (P2), (P3)>
Above, we have described a proposed model for summarizing and effectively predicting the multidimensional time series tensor X. Here, an algorithm for solving the above equation (1) will be explained. The problem here is how to determine the number of regimes and segments. This prediction system is based on the concept of minimum description length (MDL) and introduces a coding scheme that serves as a standard for generating an appropriate model.

1. Model selection and data compression Intuitively, the goodness of a model given data can be expressed by the following equation (3).

Here, Cost _M (M) indicates the model cost for expressing the model M, and Cost _C (X|M) indicates the encoding cost of the tensor X when the model M is given. α is the weight for the encoding cost (by default, α = 1), and a larger value of α produces a model that is more accurate to the real data (i.e., the number of segments m and the number of regimes r are larger) ).

・Model cost Specifically, the representation cost of the entire parameter set of this prediction system is composed of the following elements.

Note that log ^* shown in ^*2 above represents the universal code length of an integer, and log ^* (x)≈log ₂ (x) + log ₂ log ₂ (x) +... (Non-Patent Document 23). Also, if the floating point cost is c _F , then a single regime parameter θ with k states is Cost _M (θ)=log ^* (k)+c _F (k+k ² +2kd), regime The transition matrix Δ requires a cost of Cost _M (Δ)=c _F r ² .

・Encoding cost The encoding cost of X when model parameters are given can be expressed as follows (Equation 6) using negative log likelihood by information compression using Huffman codes. .

Here, the i and (i-1)th segments belong to the u and vth regimes, respectively, and X[s _i ] represents a partial sequence composed of segments s _i included in X. . P(X[s _i ]|θ _u ) is the likelihood of X[s _i ] when θ _u is given. In conclusion, the proposed algorithm determines the number r of time series patterns included in X and the number m of their changing points so as to minimize the above equation (3).

Next, a specific algorithm for realizing long-term label prediction will be detailed while summarizing data based on a cost function.

2. Algorithm Overview This prediction system consists of the following algorithm.

・REGIMEGENRATION (P1): Detects the type and change points of time series patterns included in tensor X. The dynamics of each time-series pattern is expressed as model parameters Θ, and a model parameter set {m, r, S, Θ, F} is obtained.

- FEATUREEXTRACTION (P2): Using the time series pattern summary information {m, r, S, Θ, F}, the original tensor X is expressed as a latent state tensor Z and an error tensor ε.

・SPLITCAST (P3): Features that indicate a failure from the partial sequence {Z(t _s : _{t e )} ,ε(t _s :t _e )} of a certain window t _s :t e in {Z, ε} , and predict the failure label y(t _e +l _s ) ahead of l _s .

Figure 3 shows an overview of the proposed model. When a tensor X is given, the proposed method captures the time transition of the time series pattern of X and the equipment-specific pattern, and summarizes X as {Z, ε} based on that. Finally, the alert label for l _s steps ahead is predicted and output from the obtained {Z, ε}.

3. Regime Generation (P1)
Here we describe the details of the algorithm. A fundamental question in time series analysis is whether there is a hidden structure inherent in time series data. The multidimensional time series tensor X handled here has characteristics from multiple viewpoints. That is, they are characteristics of the time domain and characteristics of the equipment domain. Specifically, time-series sensor data obtained from smart factories has time transition patterns for each process and equipment-specific patterns. Therefore, in the following, we simultaneously perform grouping and multifaceted pattern discovery that concisely summarizes the underlying structure of a given time series tensor.

Here, we propose V-Split and H-Split, which are algorithms for multilateral analysis of time series tensors. V-Split estimates the regime from the perspective of time, and H-Split expresses the characteristics of each facility as a regime. By performing these two algorithms in any direction, important patterns can be efficiently and effectively discovered from multiple angles and summarized as a regime. Specifically, the following two algorithms are repeated based on equation (3).

・V-Split: Detects time-transition patterns and their changing points from tensor X, and divides it into two groups (i.e., regimes). For these two regimes, model parameters {θ ₁ , θ ₂ , Δ} are estimated.

・H-Split: Extracts the features of each equipment from one regime that appears in tensor X, divides it into two regimes, and then estimates the model parameters of those regimes.

Using the above algorithm, the number of regimes changes as r = 1, 2, . . . If, when regime θ ₀ is divided into two regimes {θ ₁ , θ ₂ }, the value of the cost function (formula (3)) becomes large, θ ₀ is considered optimal and is not divided any further. The cost calculation is repeated in the same way for all generated regimes, and the above division algorithm is repeated until the cost no longer decreases. Finally, the segment, regime, and model parameters {m, r, S, Θ, F} are output when the cost converges, and RegimeGeneration ends.

Next, the splitting algorithms V-Split and H-Split will be described.

(3-1)V-Split
When a multidimensional time series tensor X is given, V-Split detects two regimes from the perspective of time transition and estimates their model parameters {θ ₁ , θ ₂ , Δ}. In order to generate a highly accurate model, this prediction system repeatedly detects segments/regimes and updates model parameters as follows.

- (Phase 1) V-Assignment: When two model parameters are given, two segment sets {S ₁ , S ₂ } and pattern change points are extracted based on them.

- (Phase 2) ModelEstimation: When two segment sets are given, model parameters {θ ₁ , θ ₂ , Δ} are updated based on them.

An overview of V-Split is shown in Algorithm 1 (Table 2). Algorithm 1 above is based on the expectation maximization method (EM), and each phase corresponds to E and M steps.

First, consider the case where a tensor X and two model parameters {θ ₁ , θ ₂ , Δ} are given as the simplest subproblem. V-Assignment can detect changing points in the pattern of X based on the model parameters of the regime (Steps 5 to 7 in Table 2). The transition diagram of FIG. 4 is shown to explain the basic concept of the proposed algorithm. The pattern transition between the given regimes is estimated by connecting the transitions of the two regimes {θ ₁ , θ ₂ } and comparing the encoding costs of the two regimes at each time. This algorithm calculates the encoding cost Cost _T (X|Θ)=-ln P(X|Θ) based on the Viterbi algorithm (Non-Patent Document 7), which is a type of dynamic programming. Specifically, the likelihood P(X|Θ) is calculated as shown below (Equation 7).

Here, P(X|Θ) _i indicates the likelihood of transitioning to the i-th regime θ _i . As an example, P(X|Θ) ₁ is calculated as follows (Equation 8).

Here, p _1;i (t) represents the maximum probability of latent state i in regime θ ₁ at time t, δ ₂₁ is the regime transition probability from regime θ ₁ to θ ₂ , max _u {p _{2 ;u} (t-1)} is the probability of being a plausible latent state of θ ₂ at the previous time t-1, π _1;i is the initial probability of latent state i of θ ₁ , b _1;i (x (t)) represents the output probability of x(t) for latent state i of θ ₁ , and a _1;ji represents the transition probability from latent state i to latent state j of θ ₁ . Here, the probability of regime θ ₁ at time t=1 is given by p _1;i (1)=δ ₁₁ π _1;i b _1;i (x(t)). Note that the BaumWelch algorithm (non-patent document 1) is used to estimate the model parameters, and the regime transition probability Δ={δ ₁₁ , δ ₁₂ , δ ₂₁ , δ ₂₂ } is calculated as shown in the following (Equation 9). .

Here, Σ _s∈S1 |s| represents the total length of segments belonging to regime θ ₁ , and N ₁₂ represents the number of times the regime is switched from θ ₁ to θ ₂ . δ ₂₁ and δ ₂₂ can also be calculated in the same way.

(3-2)H-Split
So far, we have explained Algorithm 1, V-Split, for capturing temporal features from the time series tensor X. As a practical matter, the time series tensor X has not only the time transition of the pattern but also individual differences between each piece of equipment. For example, even when two pieces of equipment process the same part, individual differences occur in the behavior of sensor data between the pieces of equipment for each process. In this prediction system, we propose H-Split, an algorithm that captures these unique characteristics of equipment and effectively models them. Intuitively, like V-Split, Algorithm 2 estimates an appropriate regime and its model parameters by repeating two phases: (Phase 1) regime splitting and (Phase 2) model estimation. do. What differs from V-Split is the algorithm H-Assignment (Phase 1), which captures equipment-specific characteristics. An overview of H-Assignment is shown in Algorithm 2 (Table 3). The algorithm shown in (Table 3) corresponds to the part corresponding to "V-Assignment" in step 5 in (Table 2), and H-Split is the content replaced with H-Assignment (Table 2). All you have to do is execute.

Unlike traditional clustering algorithms, H-Assignment effectively extracts equipment-specific patterns. Specifically _, given _the tensor Assign the segment of equipment i to the regime with the lower cost.

Here, X[i]={s ₁ , s ₂ ,. . . } is the set of segments of equipment i. That is, segments of the same equipment are constrained to belong to the same regime.

4.FeatureExtraction(P2)
So far, we have described an algorithm for multifacetedly detecting time series patterns that change at arbitrary timing from a multidimensional time series tensor. Next, in order to achieve long-term prediction of failure occurrence, it is necessary to extract features that indicate the cause or sign of failure from time-series data. Generally, sensor data collected at a high sampling rate contains a lot of noise, and the more complex the system being monitored, the more difficult it is to model its accurate behavior. Therefore, in this prediction system, we propose a method to abstract X using the characteristics of time-series patterns and effectively extract signs of failure. Specifically, given a time series tensor X and a model parameter set {m, r, S, Θ, F}, the error tensor when X is modeled as a latent state tensor Z based on a time series pattern Divide into ε.

_Now , assuming that r regime sets _Θ ={ _{θ 1 ,} ^. ₌₁ is transformed to any state z _i (t) of the regime in Θ. Here, z _i (t) represents the mean and variance set {μ,σ} of all data points belonging to the same state as itself. That is, the dimension of the latent state tensor is Z∈R ^w×2d×n . Next, when Θ is given, the encoding error of the measurement value x _ij (t)∈X of sensor j of equipment i at time t is expressed as a posterior probability p(x _ij (t)|θ). That is, the encoding error of the entire time series tensor X is εεR ^w×d×n . Finally, a sequence X'∈R ^w×3d×n, which is a combination of two features, is output. Through the above processing, potential behavior in the time series direction can be taken into account when estimating a learning model without losing information on input data.

5.SPLITCAST(P3)
The ultimate goal of this prediction system is to perform long-term prediction of l _s steps ahead with high accuracy from a given time series tensor X. In recent years, many methods based on deep learning have been proposed as typical methods for label prediction tasks. Methods based on deep learning can achieve flexible learning by creating multiple intermediate layers or increasing the number of units in the intermediate layer, but as the number of layers and units increases, the number of learning parameters increases and the calculation time increases. Become. There is also the problem of overfitting, and while there are many techniques to solve this problem, they are all based on empirical rules and require extremely detailed manual tuning. Therefore, this prediction system combines a feature extraction method based on a probabilistic model and a deep learning method to learn characteristic time-series patterns extracted from real data, which enables learning with a smaller network and solves the problem of overfitting. Achieving efficient and effective alert label prediction while reducing

Specifically, in order to model the time evolution of the tensor . LSTM is a deep learning model that treats input samples as time-series data and makes it possible to learn high-dimensional nonlinear dynamics. LSTM is an RNN (Recurrent neural network) in which the middle layer unit is replaced with a special structure called a memory unit, and it uses three types of input gates, output gates, and forgetting gates to calculate the unit value c _t at time t. and the output value h _t of the unit. When the output values of each gate are respectively i _t , o _t , f _t , forward propagation of LSTM is expressed by the following equation (Equation 11).

This prediction system uses the sigmoid function as the activation function. As is well known, LSTM is capable of learning long-term dependencies of input sequences given by a memory unit, and thus remembers features that are particularly important for equipment failures in the process of regime transitions and state transitions within regimes. It is conceivable to extract a feature vector that summarizes the latest operating status of the equipment.

Finally, use h _t to predict the label for l _s steps ahead. In this embodiment, prediction of a failure in the future from the latest partial sequence at time t is handled as a two-class classification task, and the output is the probability of failure occurrence at time t+ _{l s .} Therefore, the final output of this prediction system is as shown in (Equation 12).

In addition, the objective function that the model in this prediction system should minimize is BCE (Binary cross entropy), and if the batch size during model learning is N and the output value in this prediction system for each input sample i is y ^{^} _i , It is expressed as shown in (Equation 13).

The important point here is that this prediction system uses a model with a relatively small number of units (=10) and a simple structure, but has shown extremely high performance as shown in the evaluation experiment below.

(5-1) Theoretical analysis The amount of calculation in this prediction system is linear (O(wdn)) with respect to the data size. This auxiliary (substantive) theorem will be explained below.

In each iteration, V-Assignment, H-Assignment, and ModelEstimation require O(wdnk ² ) computational complexity due to encoding cost and model parameter estimation. Here, w is the number of equipment, d is the number of dimensions, n is the length of the time series, and k is the number of hidden states in the regime {θ _i } ^r _i=1 . Therefore, the amount of calculation for RegimeGeneration (P1) is O(#iter wdnk ² ). Here, the number of iterations #iter and the number k of hidden states are very small constants and can be ignored. Therefore, the amount of calculation for RegimeGeneration is O(wdn). In FeatureExtraction (P2), the amount of calculation is O(wdn) because it outputs the latent state of each equipment, each sensor, and each time and the error when modeling. Finally, when learning the obtained model using LSTM with the number of units u, the amount of calculation becomes O(u ² wdn). Here, in this prediction system, a complicated neural network is not assumed, and the number u of units of the neural network is a very small constant and can therefore be ignored. Therefore, the amount of calculation in this prediction system is O(wdn).

<Evaluation experiment>
In order to verify the effectiveness of this prediction system, we applied the specific example in Figure 2 and conducted an experiment using actual data. In this experiment, the following items were verified.

(1) Accuracy of the proposed method for long-term prediction of equipment failures
(2) Verification experiments on calculation time for real-time monitoring of equipment were conducted on a Linux (registered trademark) (Ubuntu 18.04 LTS) machine equipped with 128 GB of memory and NVIDIA TITAN V 12 GB GPU. In addition, the data set was used after normalization (z-normalization) using the mean value and variance value.

1. Prediction accuracy of this prediction system We verified the failure prediction accuracy for a given time series tensor. As a comparative example, we use a general binary prediction model, Logistic regression (LR) (Non-Patent Document 1), a recurrent neural network model, RNN (Recurrent neural network), and GRU (Gated recurrent unit). Patent Document 4) and LSTM were adopted. In LR, the mean value, variance value, maximum value, and minimum value were calculated from the partial sequence given as a mini-batch when estimating other recurrent models, and label prediction was performed as a four-dimensional feature vector. RNN, GRU, and LSTM performed label prediction using real data as input.

Regarding this prediction system, we conducted an experiment with the number of prediction steps as 200, the window size as 400, and the encoding cost weight (α=) 1.0 as defaults. In addition, for all recurrent models including this prediction system (Proposed, Figure 5), the number of units in the intermediate layer is 10, the number of units in the output layer is 5, and Adam (Non-patent Document 12) is used as the optimization algorithm. used. Accuracy was used as the evaluation index, and the average values obtained when 5-fold cross validation was performed were compared.

The data set used was the rotational speed (Speed) installed in 55 factory equipment that was actually in operation at Mitsubishi Heavy Industries Engine & Turbocharger Corporation for three months from October 2017 and was processing bearing housings. , operating voltage (Load), and equipment temperature (Temp) at 5-second intervals. Training samples are generated using a sliding window, and samples taken when the equipment itself is not in operation are omitted. The number of samples during normal operation is 62983, and the number of samples before emergency stop is 1069, which causes bias in learning. Therefore, the number of samples during emergency stop is equal to the number of samples during normal operation, and as a result, 1069 × 2 samples are used. We conducted an experiment.

(1) Prediction accuracy when changing the number of prediction destination steps FIG. 5 is a comparison diagram of accuracy when changing the prediction destination step number l _s . In the figure, the type notation of the comparative example and the data display order (left and right) correspond. In this experiment, samples were generated for each different l _s , and learning and prediction were performed. While the comparative example shows prediction accuracy comparable to that of random prediction (Accuracy=0.5), the present prediction system shows excellent performance under all conditions. From this result, it is thought that the cause of the emergency stop is not a simple one such as a rise in temperature or a drop in operating voltage, but a complex event with nonlinearity. This prediction system can capture the dynamics at each time by taking into account the time-series patterns contained in the actual data, so it has succeeded in extracting the causes of emergency stops more effectively than other recursive models. .

(2) Prediction accuracy when changing window size FIG. 6 is a comparison diagram of prediction accuracy when changing the window width of the mini-batch used during network learning. This prediction system shows stable and high performance even for data with different window widths.

(3) Precision and recall of prediction results FIG. 7 is a diagram showing precision and recall of prediction results. Precision rate indicates the ratio of the total number of predicted events to the total number of correct events. Recall rate indicates the ratio of the number of correct values for all events to the total number of correct values among predicted events. Both approaches 1 when the accuracy is high. This prediction system shows excellent performance for both indicators.

(4) Prediction accuracy with respect to the number of discovered segments FIG. 8 is a diagram showing the prediction accuracy of this prediction system with respect to the number of detected segments m. The number of detected segments was increased or decreased while changing α, which is the weight of the encoding cost, from 0.1 to 10. As shown in FIG. 8, the prediction accuracy varies greatly depending on the number of segments divided by this prediction system. When m is small, it is not possible to obtain sufficient summary information from time-series data, resulting in a decrease in prediction accuracy. Similarly, when m is large, the prediction accuracy decreases because the summary information approaches the actual data. This result also suggests that pattern detection from time-series tensors is effective in improving failure prediction accuracy. In this experiment, the best result (Accuracy=0.88) was obtained when m=1000. In conclusion, this prediction system achieved an average accuracy improvement of about 62% compared to the comparative example.

(5) Relationship between the number of training samples and prediction accuracy In actual operation, if the number of training samples is small, sufficient accuracy may not be obtained. FIG. 9 is a diagram showing the relationship between the number of learning samples and prediction accuracy. This prediction system shows higher performance than the comparative example even with a small number of samples, and as the number of learning samples increases, it is able to predict failure events with higher accuracy.

2. Calculation speed of the proposed method FIG. 10 is a diagram showing the calculation cost of the present prediction system when the number of equipment w, the number of sensors d, and the sequence length n are varied. More specifically, it is the calculation time when the input data is divided into time-series patterns and the model training is completed for 10 epochs. This prediction system efficiently detects time-series patterns from a given time-series tensor, so the amount of calculation is linear with the data size in all experiments (i.e., O(wdn)), and it is This method was found to be suitable for analysis.

As described above, this prediction system has been tested using actual data obtained from factory equipment, for example, and has demonstrated that this prediction system can appropriately model complex time-series patterns and accurately predict long-term failures. It was confirmed that this method can be used, and it was also confirmed that a significant improvement in accuracy and performance was achieved compared to existing comparative examples.

The present invention is not limited to predicting alert events for factory equipment, but can also be used to predict alert labels such as failures depending on the driving condition of each vehicle using various on-board sensors, and predict alert labels based on various biological information. Applicable. Further, the alert label can be set with various alert contents depending on the application target, in addition to defects, failures, and quality deterioration. Furthermore, the prediction processing is not limited to artificial intelligence (AI), and other techniques may be employed.

As explained above, the event prediction system according to the present invention can estimate model parameters of multifaceted dynamic patterns from time-series sensor data continuously collected from multiple types of sensors placed on multiple observation targets. a first feature extraction means that continuously performs extraction, and a second feature extraction that sequentially converts the time-series sensor data into features using the model parameters into summary information including modeling information and its error information. It is preferable to include a prediction means for inputting the summary information and outputting a probability of occurrence of a predetermined event in a predetermined time period.

Further, in the event prediction method according to the present invention, the first feature extracting unit of the computer continuously collects data from multiple types of sensors placed on multiple observation targets, and stores the time series in the storage unit. Model parameters of a multifaceted dynamic pattern are continuously extracted from sensor data and stored in the storage unit, and a second feature extracting unit of the computer extracts the model parameters and the time-series sensor data from the sensor data. The time-series sensor data is read out from the storage unit and sequentially converted into feature quantities into summary information including modeling information and its error information, and stored in the storage unit, and the prediction unit of the computer converts the summary information into the storage unit. It is preferable that the probability of occurrence of a predetermined event in a predetermined period of time is output.

Further, the program according to the present invention continuously extracts model parameters of multifaceted dynamic patterns from time-series sensor data continuously collected from multiple types of sensors placed on multiple observation targets. a first feature extracting means; a second feature extracting means for sequentially converting the time-series sensor data into summary information including modeling information and its error information using the model parameters; Preferably, the computer functions as a prediction means that outputs as input the probability of occurrence of a predetermined event in a predetermined time period.

According to these inventions, time-series sensor data is continuously collected from multiple types of sensors placed on multiple observation targets, and model parameters of multifaceted dynamic patterns are determined from the collected time-series sensor data. Extraction is continuously performed by the first feature extraction means. Next, the time-series sensor data is sequentially converted into features by the second feature extracting means using the model parameters into summary information including modeling information and its error information. Then, the prediction means inputs the summary information and outputs the probability of occurrence of a predetermined event in a predetermined time period. Therefore, there is no need for prior knowledge of time-series patterns contained in time-series sensor data, and changing points and potential behavior of patterns (regimes) can be determined based on, for example, time transitions and multiple perspectives between observation targets. be understood. Furthermore, it becomes possible to discover characteristic patterns in large-scale time-series sensor data and use them to predict events over a long period of time. Note that the sensors may be placed directly on the observation target, or may be installed in such a manner that the observation target can be observed remotely.

Further, it is preferable that the first feature extraction means detects the dynamic pattern by segmenting the dynamic pattern in the time direction and between the observation objects and patterning the segment. According to this configuration, since dynamic patterns are extracted from multiple angles, it is possible to reduce the amount of data required for processing while suppressing a decrease in accuracy.

Further, it is preferable that the first feature extraction means sets the number of segments using a cost function. According to this configuration, in segmenting time-series sensor data, the number of segments is set to an optimal value using a cost function in consideration of data amount and processing time.

Further, it is preferable that the prediction means obtains the probability of occurrence of the predetermined event based on parameters set in a neural network model. According to this configuration, highly accurate predictions can be made using a model with a small and simple structure.

Further, it is preferable that the prediction means applies LSTM (Long-short term memory) to the neural network model. According to this configuration, LSTM can be applied to deep learning models, and long-term dependence of input sequences can be learned, making it possible to make highly accurate predictions for a long period of time.

Further, the present invention captures the summary information obtained by the second feature extracting means for a predetermined period, performs machine learning with a learning prediction means having the same configuration as the prediction means, and obtains a learning result. It is preferable to include a machine learning device that updates the predicted parameters to the prediction means. According to this configuration, it is possible to gradually improve prediction accuracy.

1 Event prediction system 11 Data import processing section 12 Feature amount extraction section (first and second feature amount extraction means)
13 Prediction unit 14 Parameter update unit 100 Storage unit 20 Observation target 21 Sensor group 30 Machine learning device

Claims (4)

Multifaceted analysis is performed from time-series sensor data that is continuously collected from multiple types of sensors placed on multiple observation targets and appears in the window of ts:te (a predetermined period from the current time te to the past direction ts). a first feature extraction means that continuously extracts a model parameter set including model parameters of a dynamic pattern;
When the time-series sensor data and the model parameter set are given, the time-series sensor data is modeled using the model parameter set as modeling information Z(ts:te) based on the time-series pattern. a second feature extraction means that sequentially converts the time-series sensor data and the error information ε(ts:te) of the modeling information Z(ts:te) into features;
a prediction means for inputting the modeling information Z(ts:te) and the error information ε(ts:te) and outputting the probability of occurrence of a predetermined event in a predetermined time ahead;
The first feature extracting means detects the dynamic pattern by segmenting and patterning the dynamic pattern in the time direction and between the observation targets, and sets the number of the segments using a cost function,
The model parameter set is {m, r, S, Θ, F},
The second feature extracting means uses a hidden Markov model to summarize the modeling information Z(ts:te) and error information ε(ts:te) during modeling,
The event prediction system is characterized in that the prediction means obtains the probability of occurrence of the predetermined event based on parameters set in a neural network model applying LSTM (Long-short term memory).
However, m is the number of segments in the time-series sensor data, r is the number of regimes among them, S is the segment set representing the start point, end point, and observation target number of each segment, Θ is the number of each regime. model parameters, and F is the number of the regime to which the segment belongs. A learning prediction having the same configuration as the prediction means takes in the modeling information Z(ts:te) and the error information ε(ts:te) obtained by the second feature extracting means for a predetermined period. 2. The event prediction system according to claim 1 , further comprising a machine learning device that performs machine learning in said prediction means and updates said prediction means with parameters obtained as a learning result . The first feature extracting means of the computer continuously collects data from a plurality of types of sensors placed on a plurality of observation targets, and ts:te (a predetermined period from the current time te to the past direction ts). We continuously extract model parameter sets including model parameters of multifaceted dynamic patterns from time-series sensor data appearing in the window of
When the second feature extracting means of the computer is given the time-series sensor data and the model parameter set, the second feature extracting means of the computer converts the time-series sensor data into modeling information Z based on the time-series pattern using the model parameter set. (ts:te) and error information ε(ts:te) between the time-series sensor data when modeled and the modeling information Z(ts:te), and sequentially convert them into features,
A prediction means of the computer inputs the modeling information Z(ts:te) and error information ε(ts:te) and outputs the probability of occurrence of a predetermined event in a predetermined time ahead, and further,
The first feature extracting means detects the dynamic pattern by segmenting and patterning the dynamic pattern in the time direction and between the observation targets, and sets the number of the segments using a cost function,
The model parameter set is {m, r, S, Θ, F},
The second feature extracting means uses a hidden Markov model to summarize the modeling information Z(ts:te) and error information ε(ts:te) during modeling,
The event prediction method is characterized in that the prediction means obtains the probability of occurrence of the predetermined event based on parameters set in a neural network model applying LSTM (Long-short term memory).
However, m is the number of segments in the time-series sensor data, r is the number of regimes among them, S is the segment set representing the start point, end point, and observation target number of each segment, Θ is the number of each regime. model parameters, and F is the number of the regime to which the segment belongs. Multifaceted analysis is performed from time-series sensor data that is continuously collected from multiple types of sensors placed on multiple observation targets and appears in the window of ts:te (a predetermined period from the current time te to the past direction ts). a first feature extraction means that continuously extracts a model parameter set including model parameters of a dynamic pattern;
When the time-series sensor data and the model parameter set are given, the time-series sensor data is modeled using the model parameter set as modeling information Z(ts:te) based on the time-series pattern. a second feature extraction means that sequentially converts the time-series sensor data and error information ε(ts:te) between the time-series sensor data and the modeling information Z(ts:te) into features;
The computer functions as a prediction means that inputs the modeling information Z (ts: te) and error information ε (ts: te) and outputs the probability of occurrence of a predetermined event in a predetermined time ahead, and further,
The first feature extracting means detects the dynamic pattern by segmenting and patterning the dynamic pattern in the time direction and between the observation targets, and sets the number of the segments using a cost function,
The model parameter set is {m, r, S, Θ, F},
The second feature extracting means uses a hidden Markov model to summarize the modeling information Z(ts:te) and error information ε(ts:te) during modeling,
The prediction means is a program that outputs an occurrence probability of the predetermined event based on parameters set in a neural network model to which LSTM (Long-short term memory) is applied.
However, m is the number of segments in the time-series sensor data, r is the number of regimes among them, S is the segment set representing the start point, end point, and observation target number of each segment, Θ is the number of each regime. model parameters, and F is the number of the regime to which the segment belongs.

Images